AI is everywhere but in the productivity figures

This year, Alphabet, Amazon, Microsoft and Meta plan to spend over $700 billion, much of it on AI.¹ However, the productivity figures tell a mixed story. US productivity grew by 2.2% over the past year, but only one in five American firms say they use AI.² In the UK and euro area, productivity grew by only 0.7% and 0.6% respectively.³ The technology is said to bring about the biggest revolution since electricity; these figures say something different. The question is whether this is an anomaly or the norm for a new technology, so by looking back at steam, electricity and computers, we can try to judge whether AI is different.

Steam

Watt patented his improved steam engine in 1769. It used less than half the coal of its predecessor, the Newcomen engine, but it ran at low pressure. As a result, steam spread slowly: it only matched water as a power source in 1830, and even in 1870 nearly half of it was used in mining and cotton. High-pressure engines reduced coal use by a further 60%, yet they required much-improved boilers and only became practical in textile mills from the 1840s onwards.⁴ According to economic historian Nicholas Crafts’ estimates, steam, therefore, had little effect on productivity growth before 1830, and its impact peaked between 1850 and 1870, about 80 years after Watt’s patent. For this revolution, the problem was the technology itself. Steam only paid off once it had been improved enough to be worth adopting widely – long after its initial development.

So, is AI being held back for the same reason? Is the gap between the technology and the productivity a symptom of lacklustre AI capabilities? The evidence suggests not. Researchers at Epoch AI found that AI capabilities have recently accelerated rather than slowed, mainly in programming and maths.⁵ If the technology is not the bottleneck, then the problem must lie elsewhere.

Steam took decades to pay off

Electricity

Electricity tells a similar, but importantly distinct story. Edison opened his first central power stations in the early 1880s, yet even in 1899 electric motors supplied less than 5% of the mechanical power in US factories and only around half by the early 1920s.⁶ As shown by the economic historian Paul David, contrary to the problem faced in steam adoption, the issue this time was how the technology was used. Initially, factories replaced their steam engines with electric motors, but they kept their old systems of shafts and belts, so productivity was barely affected. The true gains from electricity only came when factories redesigned their whole system of production around a small motor on each machine. Once this happened in the 1920s, productivity growth in US manufacturing rose from 0.2% a year between 1909 and 1919 to over 5% a year between 1919 and 1929.⁷ Compounded over a decade, the same capital and labour were producing over 60% more output in 1929 than in 1919.

In this case, the problem was not the machine itself but the way factories were organised around it, and AI may face the same issue. A 2025 MIT study found that 95% of companies’ AI trial projects never made it beyond the trial stage, mainly because firms struggled to adapt the way they worked to incorporate AI effectively, not because the technology failed.⁸ The payments firm Klarna is a good example of this. In 2024, it claimed its AI could do the work of 700 customer service staff; one year later, its boss admitted that the AI’s service was “lower quality” and the firm started hiring people again.⁹ Like factories that fitted a motor to their old shafts and belts, firms may just be sticking AI on to existing workflows, instead of redesigning their operations around this new technology.

Computers

“

“You can see the computer age everywhere but in the productivity statistics”

Nobel Memorial Prize-winning economist Robert Solow, 1987.¹⁰

For its time, Solow’s prognosis was correct; eventually, however, the gains did arrive as US labour productivity growth rose from around 1.4% per year from 1974 to 1990, to 2.4% between 1996 and 2001. In the latter time period, information technology accounted for about three-quarters of all US productivity growth.⁴

Why the delay in this case? The three economists Erik Brynjolfsson, Daniel Rock and Chad Syverson put forward an explanation, describing their theory of a “productivity J-curve”.¹¹ To use computers optimally, firms spent significant sums of money on reorganising labour and capital, retraining workers and writing new software. Statistics do not count this spending as investment, so productivity initially looks weak while all this money is being spent, and suddenly it looks unusually strong once this invisible investment has paid off. The three economists estimate that, once adjusted for this investment, US total factor productivity at the end of 2017 was 15.9% higher than was shown by official figures.

If AI follows a similar pattern to these technologies, today’s weak figures may not mean AI is failing, but that the economy is still only at the bottom of the curve. So what would climbing the curve look like today? Firms would need to invest in the less visible side of AI: organising their data, retraining staff and redesigning jobs. One firm’s experience shows what this can look like: in a study of 5,172 customer-support agents, economists Erik Brynjolfsson, Danielle Li and Lindsey Raymond found that AI raised the number of problems solved per hour by an average of 15% with the biggest gains for less experienced workers.¹² However, these firms remain an exception: only about one in five American firms use AI at all, so most of the economy may not yet have started to climb.

Is AI different?

Despite history suggesting a long wait, there are genuinely good reasons to think AI’s productivity gains will appear faster. Each sequential technological breakthrough, so far, has delivered meaningful improvement sooner than the one before: around 80 years for steam, 40 for electricity and 25 for computers. These lags, however, are not measured from consistent starting points, so the pattern is more of a rough guide than a precise trend.

More concretely, a 2026 study by the World Intellectual Property Organization found that the gap between a technology’s invention and its adoption in a typical country has fallen from 50 years before 1950 to 15 years after 2000.¹³ AI may move faster still. Unlike electricity, which needed new power stations and a complete redesign of factories, AI runs mostly on computers and networks which are already owned by firms, supporting the case for the expectation of a faster adoption.

However, looking at past technologies also reveals an unfortunate caveat. Looking again to electricity, the factor limiting productivity growth was not the technology itself, but the reorganisation that was required in response. Impactful AI integration will likely need a similar transition in human capital: new skills, new jobs, new workforce structures – all of which take several years on an aggregate level. If firms are beginning to make these changes and investments now, we may be able to look to the J-curve theory and infer that official statistics will understate AI’s effects for quite a while. Thus, perhaps the weak figures of today are neither unexpected nor a signal that AI is failing, but are instead exactly what history seems to predict and the start of a historic productivity boom.

The experts do not agree

Nobel Prize-winning economist Daron Acemoglu estimates that AI will raise total factor productivity by no more than 0.66% over the next decade, the equivalent of adding around 0.07 percentage points a year to productivity growth.¹⁴ Global investment bank Goldman Sachs, on the other hand, expects AI to add 1.5 percentage points to productivity growth each year for a decade, more than twenty times as much.¹⁵

Forecasts of AI's gains are far apart

History and probability suggest that reality of AI’s effect lies between these two credible forecasts. It is probable that AI will follow the pattern of shortening time lags and pay off faster than steam, electricity and computers, as it can be built into existing technology. However, this does not change the fact that firms still have to reorganise how they operate around AI, which is the harder and more time-consuming task. Therefore, the gains are unlikely to be as fast as the optimists hope either.

For now, it is true that we can safely sit in Solow’s camp: AI is “everywhere but in the productivity statistics”. What we can learn from history, however, is that whether this is still true by 2036 depends as much on firms as on the technology itself.

Footnotes

1. Sherwood News, Alphabet, Amazon, Microsoft and Meta plan more than $700 billion on capex this year (opens in a new tab), 30th April 2026.

2. The Productivity Institute, Q2 2026: U.S. Productivity Growth Revival Continues (opens in a new tab), 2026.

3. Office for National Statistics, Productivity flash estimate and overview, UK: April to June 2026 and January to March 2026 (opens in a new tab), 18th August 2026.The Productivity Institute, Q1 2026: Euro Area Productivity - A Better Release Than The Headline Suggests (opens in a new tab), 2026.

4. London School of Economics, Steam as a General Purpose Technology: A Growth Accounting Perspective (opens in a new tab), May 2003. ²

5. Epoch AI, Have AI capabilities accelerated? (opens in a new tab), 16th April 2026.

6. American Economic Review, The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox (opens in a new tab), May 1990.

7. Oxford University Press for the British Academy, General Purpose Technologies and Productivity Surges: Historical Reflections on the Future of the ICT Revolution (opens in a new tab), 2003.

8. Computing, MIT report: 95% of corporate generative AI pilots fail to deliver returns (opens in a new tab), 2025.

9. Entrepreneur, Klarna Is Hiring Customer Service Agents After AI Couldn't Cut It on Calls, According to the Company's CEO (opens in a new tab), 9th May 2025.

10. The New York Times, We'd Better Watch Out (opens in a new tab), 12th July 1987.

11. American Economic Journal: Macroeconomics, The Productivity J-Curve: How Intangibles Complement General Purpose Technologies (opens in a new tab), January 2021.

12. Quarterly Journal of Economics, Generative AI at Work (opens in a new tab), May 2025.

13. World Intellectual Property Organization, How Do New Technologies Diffuse? (opens in a new tab), 2026.

14. National Bureau of Economic Research, The Simple Macroeconomics of AI (opens in a new tab), May 2024.

15. Goldman Sachs, Generative AI could raise global GDP by 7% (opens in a new tab), 5th April 2023.